Functional neuroimaging technologies, including functional magnetic resonance imaging (fMRI) and positron emission tomography (PET), are powerful noninvasive tools for mental health research. Analytic methods for fMR! and PET data are critical both for determining substantive research questions that can be addressed and for ensuring the validity of inferences. This project seeks to develop state-of-the-art statistical methodology for fMRI and PET data that will have important mental health implications for clinical practice and research. The long-term goals are 1) to assist physicians in making treatment decisions for patients with psychiatric disorders, focusing here on schizophrenia and major depression and 2) to establish a modeling framework for characterizing task-related brain activity that accounts for spatial associations arising, for example, from complex neurophysiological links between brain regions. In an effort to make an impact on clinical mental health practices, one aim is to develop a novel approach to predict individual-specific responses to treatment. Specifically, the goal is to predict post- treatment patterns of task-related brain activity for a particular patient based on pre-treatment scans and other patient characteristics and to predict eventual symptom response to treatment. The planned developments entail constructing and validating a Bayesian hierarchical model and an accurate classification algorithm for schizophrenia patients and for never-treated depressed subjects. A second aim is to construct a Bayesian hierarchical model for making inferences regarding task-related changes in brain activity, accounting for functional associations between different spatial locations (voxels). This approach would yield localized estimates, similar to commonly applied methods, but would estimate and adjust for key functional linkages. By building a model based on assumptions that are well-suited to the data, a major advantage of the proposed procedure is the ability to draw localized inferences that borrow strength from related voxels, often yielding more accurate results. A second advantage is that tests about extended anatomical regions can incorporate estimates of between-voxel correlations. Spatial modeling developments from Aim 2 may give rise to extensions to the proposed prediction framework (Aim 1). Successful development of the predictive algorithms would provide results that translate naturally to a clinical setting to help inform physicians'decisions regarding psychiatric treatments. Furthermore, the proposed spatial modeling framework would be a novel contribution to existing analytic methods for functional neuroimaging data. The focus here on fMRI and PET data related to schizophrenia, depression, and cocaine-dependence illustrates the potential applicability and relevance of the proposed methods across a range of mental health disorders.